A face anonymization generation method and system based on feature decoupling

Abstract

This invention relates to a face anonymization generation method and system based on feature decoupling, belonging to the field of computer vision technology. The method includes: acquiring a face image to be processed; inputting it into an attribute-preserving bottleneck network, filtering out identity residual information in the image by applying attribute supervision and information bottleneck constraints in the latent space, and decoding to obtain a bottleneck image that preserves facial attributes but does not contain identity details; using the bottleneck image as a conditional guiding signal, inputting it into a pre-trained conditional diffusion thinning module, and using the random denoising process of the diffusion model to recover the high-frequency texture details of the image, generating the final anonymized face image. This invention abandons the reliance on explicit identity detection in traditional methods. Through a two-stage compression and regeneration mechanism, it completely conceals identity features while preserving semantic attributes crucial to downstream tasks to the greatest extent, and significantly improves the visual realism and usability of the generated image.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology and relates to a face anonymization generation method and system based on feature decoupling, which is used to maintain the consistency and usability of non-identity attribute information in face images while removing or replacing facial identity features. Background Technology

[0002] With the rapid development of deep learning and artificial intelligence technologies, large-scale models are increasingly widely used in the real world, and training or fine-tuning these models typically requires massive amounts of data. However, the resulting data privacy issues, especially the protection of facial data privacy involving personal identity, have attracted increasing attention. Facial data contains unique biometric features and is intrinsically linked to an individual's privacy and identity. With the rapid advancement of facial recognition technology, the ability to identify an individual based on a single image has raised significant privacy concerns.

[0003] To mitigate privacy risks, face anonymization technology has emerged, aiming to remove identity-specific information from images while preserving non-identity attributes. Ideally, anonymized images should still be usable for identity-neutral face analysis tasks, such as face detection, expression recognition, and attribute classification. However, in practical applications, there is often a significant conflict between the strength of privacy protection and visual usability. Achieving an effective balance between the two has long been a technical challenge in the field of face anonymization.

[0004] Traditional face anonymization methods typically employ image obfuscation techniques, such as blurring, pixelation, or masking facial regions. While these methods are simple to implement and do not require complex models, they often severely degrade the visual quality of the image, losing most of the semantic information crucial for downstream tasks (such as facial expressions and gaze direction), leading to a significant reduction in data usability.

[0005] With the development of Generative Adversarial Networks (GANs), a number of deep learning-based anonymization methods have emerged. These methods (such as DeepPrivacy and CIAGAN) typically rely on explicit identity detection or prior conditions, such as pre-extracting facial landmarks, semantic segmentation maps, or latent identity vectors, and then using a generative model to replace or modify the identity portion. However, the above-mentioned schemes generally suffer from the following shortcomings: (1) High dependence on detection accuracy: Such methods rely heavily on the accuracy of identity detectors (such as key point detectors). Under extreme poses, occlusions, or complex lighting conditions, once the detector fails, the anonymization effect will be greatly reduced, and even serious visual artifacts may occur.

[0006] (2) Incomplete identity decoupling: Many existing methods attempt to separate identity features from other features in the latent space, but such separation is often difficult to achieve completely. If identity information cannot be strictly removed, attackers may still reconstruct the identity or carry out re-identification attacks through the remaining features.

[0007] (3) Limited ability to control attributes: Existing methods often struggle to precisely preserve or control specific facial attributes (such as changing only the identity but strictly preserving subtle facial expression changes) while strictly removing the identity, resulting in limited utility of the generated image in downstream tasks.

[0008] In summary, existing technologies still struggle to simultaneously meet the requirements of strong identity privacy elimination, good visual usability, and low reliance on identity detection or supervision in achieving facial anonymization. Therefore, a new technical solution is urgently needed that, starting from the feature representation and generation mechanism level, effectively weakens the correlation between facial identity and reality without explicitly modeling identity information, while retaining semantic attribute information valuable for downstream tasks, thereby improving the practicality and security of facial anonymization technology. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide a generative face anonymization technology that can effectively weaken the correlation between face identity and retain the semantic attribute information of face, overcoming the shortcomings of existing face anonymization methods in terms of privacy protection strength, visual usability, and dependence on identity detection or identity supervision.

[0010] Specifically, this invention aims to solve the following technical problems: how to suppress the expression of distinguishing information related to individual identity from the feature representation and generation mechanism level without explicitly extracting, modeling or manipulating facial identity features, thereby reducing the risk of anonymization results being re-identified or reversed; at the same time, how to maintain facial expressions, postures, structures and other identity-independent visual attributes as much as possible during the anonymization process, so as to meet the usage requirements of downstream tasks that are not related to identity, such as face detection, behavior analysis, and expression recognition.

[0011] To achieve the above objectives, the present invention provides the following technical solution: Solution 1: A face anonymization generation method based on feature decoupling. From the perspective of information constraints and generative reconstruction, this method constructs an intermediate representation that is highly expressive of facial attribute information but limited in its ability to express features unrelated to those attributes. This makes it difficult to stably preserve fine-grained features highly correlated with individual identity within this representation. Based on this, a generative reconstruction mechanism based on the intermediate representation is introduced to generate anonymized face images with natural appearance and attribute consistency. This improves anonymization strength while maintaining visual quality and application usability. This method achieves a face anonymization approach with strong privacy protection, good generalization performance, and practical value without relying on identity recognition models or requiring identity labeling or supervision information. It provides a new technical means for the secure sharing and compliant application of face data.

[0012] The method specifically includes the following steps: S1: Obtain the original face image to be anonymized.

[0013] S2: Construct and utilize the attribute-preserving bottleneck network (Stage I) to process the original face image.

[0014] An information-constrained latent space is constructed, and the original face image is input into an attribute-preserving bottleneck network. By applying attribute supervision and information bottleneck constraints to the latent space, the identity residual information in the image is filtered out, and the bottleneck image that preserves facial attributes but does not contain identity details is decoded.

[0015] S3: Construct and utilize the Conditional Diffusion Refinement Module (Stage II) to process the bottleneck image. The bottleneck image is used as a conditional guiding signal and input into the pre-trained Conditional Diffusion Refinement Module. Utilizing the random sampling generation mechanism of the diffusion model, the high-frequency texture and visual details of the image are restored to generate the final anonymized face image.

[0016] Furthermore, in step S2, the attribute-preserving bottleneck network is based on a variational autoencoder. The architecture is constructed, and its training and inference logic includes: (1) Encode the input image into a low-dimensional latent representation.

[0017] (2) Attribute decoupling mechanism: Establish a one-to-one correspondence between each dimension of the potential representation and a predefined set of facial attributes (such as gender, age, expression, etc.). Supervised training of the potential representation is performed through an attribute classifier to ensure that each dimension encodes only specific attribute information.

[0018] (3) Identity filtering mechanism: KL divergence (KL divergence) is introduced as a regularization term to limit the information capacity of potential channels. By adjusting the regularization strength, the network is forced to encode only supervised attribute information, while identity texture features not included in the attribute set are discarded as redundant information.

[0019] (4) The decoder reconstructs a bottleneck image based on the latent representation, which retains the facial structure and attributes but blurs the identity features.

[0020] Furthermore, in step S3, the conditional diffusion refinement module is constructed based on the Conditional Diffusion Probabilistic Model, and its specific mechanism includes: (1) Conditional guidance: The bottleneck image generated by the attribute-preserving bottleneck network is used as a condition signal and input into the denoising network of the diffusion model; (2) Random generation: Using the inverse denoising process of the diffusion model, the image is gradually restored starting from pure Gaussian noise and guided by the bottleneck image. Since nondeterministic identity information has been removed from the bottleneck image, the diffusion model randomly fills in reasonable high-frequency details (such as hair strands and skin pores) by learning the statistical distribution of the data. This stochasticity further blocks the possibility of recovering the original identity from the generated image.

[0021] Furthermore, regarding the attribute editing function: Since the latent representation is interpretable, users can directly modify the dimension values ​​corresponding to specific attributes in the latent representation (e.g., change the value of the "smile" attribute). The modified latent representation is decoded into a new bottleneck image and input into the conditional diffusion refinement module to generate anonymized face images with changes in target attributes, thereby achieving precise control over face attributes or balancing of dataset distribution.

[0022] Furthermore, the total loss function for training the attribute-preserving bottleneck network. for:

[0023] in, To reconstruct the loss, the L1 distance is used; For binary cross-entropy loss, The weighting coefficients for attribute supervision loss are used to adjust the proportion of attribute classification in the total loss. Let KL divergence be the KL divergence. is the regularization weight coefficient of the KL divergence, used to control the constraint strength of the information bottleneck to adjust the degree of decoupling in the latent space.

[0024] Furthermore, the denoising training loss function of the diffusion model Represented as:

[0025] in, Indicates the original image The time step of random sampling and Gaussian noise Find the expected value of the joint distribution; Indicates random noise. This represents the noise component predicted by the denoising network. For noisy images, For time steps, For conditional images.

[0026] Option 2: A face anonymization generation system based on feature decoupling, comprising: The feature extraction module is used to acquire the original face image and extract its multi-dimensional features; The feature decoupling and replacement module decouples multi-dimensional features into independent identity features and non-identity features through the attribute-preserving bottleneck network, filters out identity residual information in the image, and decodes to obtain a bottleneck image that preserves facial attributes but does not contain identity details; The image reconstruction module, through the conditional diffusion refinement module, generates anonymized face images based on the bottleneck image combination.

[0027] Option 3: An electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in Option 1.

[0028] Option 4: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in Option 1.

[0029] The beneficial effects of this invention are as follows: 1) Achieving high-strength facial identity privacy protection: This invention constructs a bottleneck representation with sufficient attributes and limited expressive power, making it difficult to stably encode individual identity distinguishing features that are unrelated to facial semantic attributes at the feature level. Furthermore, it introduces a random reconstruction mechanism in the subsequent generation process, thereby significantly weakening the correlation between facial images and original identities and reducing the risk of anonymization results being re-identified or reversed.

[0030] 2) Reduced dependence on identity recognition models or identity supervision information: In the anonymization process, this invention does not require explicit extraction, modeling or manipulation of facial identity features, nor does it rely on identity recognition models or identity annotation data. Instead, it weakens the identity correlation by constraining the feature representation capability, thereby improving the generalization ability of the technical solution in different datasets and application scenarios.

[0031] 3) Supports controllable attribute adjustment, enhancing application flexibility: Based on a structured attribute feature representation, this invention can controllably adjust the semantic attributes of faces without introducing identity information, thereby supporting flexible configuration of the attribute distribution of anonymized results. It is suitable for anonymized data construction, data balancing processing, and various practical application scenarios, improving the scalability and practical value of the overall technical solution.

[0032] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the face anonymization generation method based on feature decoupling according to the present invention; Figure 2 This is a graph showing the qualitative analysis results of the present invention and other anonymization methods on the CelebA test set; Figure 3 To indicate the accuracy of attribute classification in this invention, each attribute is represented by an abbreviation, such as "DC" representing "double chin"; Figure 4 This is a diagram showing the result of editing the attributes of this invention. Detailed Implementation

[0034] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0035] Please see Figures 1-4 This invention provides a face anonymization generation method based on feature decoupling. This method abandons the traditional detection and modification of explicit identity features (such as key points), instead filtering identity information by constructing a sufficiently attributed bottleneck representation, and then using a diffusion model to recover visual details. The method mainly includes the following steps: Step 1: Data Acquisition and Preprocessing Obtain a training dataset containing face images and their corresponding attribute labels (e.g., CelebA or CelebA-HQ dataset). Perform center alignment and size normalization on the images (e.g., resize to...). (pixels). Attribute tag collection Includes predefined, non-identity-sensitive attributes (such as "gender", "age", "whether wearing glasses", "smile", etc.), denoted as .

[0036] Step 2: Construct and train the attribute-preserving bottleneck network (Stage I) Construct a variational autoencoder ( This network aims to learn a compact latent representation. This allows it to retain information crucial for reconstructing facial attributes while filtering out identity information through strict information bottlenecks. Specifically, this includes: 1) Encoding stage: Let the input face image be... An encoder is used to encode a face image to obtain the corresponding latent feature representation. Latent feature representation The conditional distribution of a random variable confined to a low-dimensional continuous vector space can be expressed as:

[0037] in, and Let these represent the mean vector and standard deviation vector output by the encoding network, respectively. These are the learnable parameters of the encoder; The approximate posterior distribution represents the distribution given the input image. In this case, the latent variables inferred by the encoder The probability distribution; It is a multivariate Gaussian distribution, specifically referring to... As the mean vector, with The distribution of the covariance matrix; Construct operators for diagonal matrices.

[0038] 2) Attribute decoupling and supervision: To give the latent space interpretability, latent representations are forced to be decoupled. Each dimension With a specific semantic attribute Related. Among them, Indicates the first Semantic attributes of a face that are independent of identity. Latent feature representation. Constructed by It consists of several characteristic components, namely:

[0039] And constrain each feature component Used only to express the corresponding semantic attributes This results in attribute-level decoupled representations in the feature space. To this end, a set of attribute classifiers is appended to the latent space, supervised by binary cross-entropy loss (BCE).

[0040] in, It is the Sigmoid activation function. This represents the total number of attribute categories.

[0041] 3) Information Bottleneck Constraint: To filter out the remaining information (i.e., identity information) that is not defined by attributes, Kullback-Leibler (KL) divergence is introduced as a regularization term to limit the capacity of the latent space.

[0042] in, Let KL divergence be the KL divergence. It follows a standard normal distribution. This can be achieved by adjusting the hyperparameters. Controlling the tightness of the bottleneck ensures that the remaining information is only sufficient to reconstruct attributes, but not enough to reconstruct accurate identity textures.

[0043] 4) Decoding and Reconstruction: Using a decoder to extract the latent code from the samples. The image is reconstructed to obtain the bottleneck image. Reconstruction loss is calculated using the L1 distance:

[0044] in, For approximate posterior distribution Find the expected value. It is the L1 norm, representing the sum of the absolute values ​​of the vector elements.

[0045] The total loss function for Stage I is:

[0046] in, These are the weighting coefficients for the attribute supervision loss, used to adjust the proportion of attribute classification in the total loss. is the regularization weight coefficient of the KL divergence, used to control the constraint strength of the information bottleneck to adjust the degree of decoupling in the latent space. Control the strength of attribute supervision. Adjusting the density of the information bottleneck. After training, the network is able to generate blurred bottleneck images that remove identity details but retain facial structure and attributes.

[0047] Step 3: Build and train the conditional diffusion refinement module (Stage II) Due to the bottleneck image generated in stage one Often too smooth and lacking in high-frequency details, this step introduces a Conditional Diffusion Model to refine it in order to restore high-fidelity visual effects.

[0048] 1) Conditional mechanism: The bottleneck image generated in stage one As a conditional input, it guides the inverse denoising process of the diffusion model.

[0049] 2) Diffusion process: In the forward diffusion process of the generative model, the model is gradually diffused towards the real face image. Noise is injected into the image to construct a series of noisy images. Its conditional probability distribution is expressed as:

[0050] in, Indicates the diffusion time step. The preset total number of diffusion steps, For noise scheduling parameters, It is an identity matrix.

[0051] During the reverse generation process, a denoising network is used to progressively reconstruct the noisy image. The denoising network simultaneously receives the current noisy image at each time step. Time step index and the intermediate face representation image As a conditional input, the corresponding noise component is predicted, and its generation process is represented by the conditional probability distribution as follows:

[0052] in, This represents the learnable parameters of the denoising network. It is a conditional mean function. This is the variance parameter related to the time step.

[0053] 3) Training objective: The denoising network receives noisy images. Time step and conditional images As input, the generative model is optimized by minimizing the difference between the predicted noise and the real noise. Its denoising training loss function is expressed as:

[0054] in, Indicates the original image The time step of random sampling and Gaussian noise Find the expected value of the joint distribution. Indicates random noise. This represents the noise component predicted by the denoising network. It is an L2 norm.

[0055] 4) Identity blocking principle: Due to conditional images The deterministic identity clues have been removed, and the diffusion model can only generate high-frequency details (such as skin texture and hair details) randomly based on statistical patterns. This randomness further undermines the consistency of identity and prevents the leakage of identity information.

[0056] Step 4: Editing facial attributes (optional example) In some optional embodiments of the present invention, based on the attribute-preserving feature bottleneck representation constructed in the first stage, the present invention also supports controllable adjustment of facial semantic attributes during the anonymization generation process, thereby achieving flexible control over the attribute distribution of the generated result without introducing identity information.

[0057] Specifically, as mentioned earlier, the feature bottleneck representation obtained in the first stage It is constructed as a feature component consisting of multiple attributes, that is:

[0058] Each feature component With preset facial semantic attributes One-to-one correspondence. Based on this structured feature representation, while keeping other feature components unchanged, the feature component corresponding to any target attribute can be adjusted, thereby achieving controllable changes to the attribute.

[0059] In one implementation, targeting attributes By its corresponding feature components By applying the adjustment operator, the adjusted feature representation is obtained. Its form is expressed as:

[0060] in, This represents an attribute adjustment function, used to adjust the feature components corresponding to a target attribute. The attribute adjustment function can take the form of linear transformation, threshold constraint, sign reversal, or interval mapping, to adapt to the adjustment requirements of different attribute types.

[0061] Step 5: Perform face anonymization inference During the inference phase, for the face image to be anonymized: 1) Input it into the trained attribute-boosting network to extract the latent encoding. And decode it to obtain a bottleneck image without identity information. .

[0062] 2) The diffusion model, trained as conditional input, starts with a purely noisy image and then... The reverse denoising sampling step generates the final anonymized image. .

[0063] 3) Attribute editing (optional implementation): Due to potential space It is interpretable, and users can directly modify it. The corresponding dimension values ​​(e.g., setting the dimension corresponding to "smile" to 1) are then decoded and diffused to generate the data, thereby enabling the editing or balancing of specific facial attributes while anonymizing the data.

[0064] To verify the effectiveness of the face anonymization method for usability preservation proposed in this invention, the inventors conducted quantitative and qualitative experiments on three widely used face datasets (CelebA, CelebA-HQ, FFHQ) and compared it with existing mainstream anonymization methods (including DeepPrivacy, CIAGAN, FALCO, G2Face, FAMS, etc.).

[0065] The embodiments of this invention are trained on a single NVIDIA RTX 4090 GPU. In Stage I, training is performed using... Architecture, potential dimensions set as This corresponds to the number of facial attributes used in supervised training. The hyperparameters are set to... and In Stage II, a U-Net-based conditional diffusion model is used, initialized with pre-trained weights, to refine the bottleneck image generated in Stage I, and trained for 500 epochs.

[0066] Privacy protection is the core objective of anonymization. This experiment uses the identity re-identification rate (Re-ID) and average identity distance (ID-DIS) as indicators. Lower Re-ID and higher ID-DIS indicate more thorough anonymization and greater difficulty in identity recognition. The experimental results, shown in Table 1, demonstrate that the method of this invention achieves the lowest Re-ID rate on the CelebA test set, at only [value missing]. Much lower than FALCO ( ) and DeepPrivacy ( Meanwhile, this invention achieved the highest ID-DIS score (). The result indicates that the generated face is furthest from the original face in the identity feature space. This proves that the strategy of filtering identity information through the "attribute preservation bottleneck" is more effective than the traditional "detect-replace" strategy in blocking identity links.

[0067] Table 1. Privacy protection among different anonymization methods on the CelebA and FFHQ test sets.

[0068] In addition to quantitative indicators, this invention also qualitatively analyzes the visual effects of the generated images. The images generated by this invention are compared with those generated by DeepPrivacy, CIAGAN, FALCO, and other methods. Experimental results are as follows: Figure 2 The results show that existing methods (such as DeepPrivacy) typically employ a "detect-replace" strategy, replacing only the facial region. This often results in artifacts or unnatural stitching at the point where the generated face blends with the background or hairstyle. While methods based on latent spatial operations, such as FALCO, preserve the background, they often rely too heavily on the latent encoding of the original image, resulting in the generated face still retaining visual cues of the original identity (such as eye shape and unique texture).

[0069] In contrast, this invention, due to its rigorous information bottleneck filtering in Stage I, generates images that retain the semantic attributes of the original image (such as "smiling," "blonde hair," and "wearing glasses") while reconstructing the background, hair details, and other aspects. Although this reconstruction differs from the original image at the pixel level (resulting in a lower SSIM value, which is actually an advantage in anonymization tasks), the generated face has extremely high overall naturalness, with no stitching artifacts, and completely eliminates the visual features of the original identity, achieving a high-level anonymization effect that is "similar in form but not in spirit."

[0070] To verify whether the anonymized images still retain data value, this experiment evaluated image quality (NIQE, PIQE, BRISQUE) and downstream task performance (face detection rate, expression recognition, attribute classification accuracy). Experimental data are shown in Table 2 and... Figure 3 show: (1) Image quality: The images generated by this invention exhibit excellent performance in terms of naturalness and perceived quality, with an NIQE score of [insert score here]. (The lower the better), superior to CIAGAN ( ) and FALCO ( This indicates that the diffusion model effectively recovered high-frequency details.

[0071] (2) Face detection: The detection rate of anonymized images using the MTCNN detector reached [percentage missing]. This proves that the facial structure was perfectly preserved.

[0072] (3) Attribute consistency: Although the identity is removed, the generated image is highly consistent with the original image in terms of attributes such as expression and posture, which can support analysis tasks such as expression recognition and attribute classification.

[0073] Table 2. Performance comparison of different anonymization methods on the CelebA test set.

[0074] In Table 2, the bolded parts indicate the best performance, while the underlined parts indicate the second-best performance.

[0075] To verify the necessity of each module in this invention, ablation experiments were conducted, and the results are shown in Table 3: (1) Remove attribute supervision (w / o) If the attribute loss function of stage one is removed: The re-identification rate (Re-ID) has decreased to However, the attribute classification accuracy (ACC) dropped significantly to This indicates that the image has lost key semantic information, and its data value (utility) is unavailable.

[0076] (2) Remove information bottlenecks (w / o) If the KL divergence constraint of stage one is removed: The model's re-identification rate (Re-ID) increased to This indicates an increase in the leakage of identity information.

[0077] (3) Hyperparameter analysis: When At that time, the model is in privacy protection (Re-ID) ) and attribute preservation (ACC) The optimal balance was achieved between ( ).

[0078] Table 3 Ablation studies on the CelebA test set

[0079] In summary, comparative and ablation experiments demonstrate that this invention, by combining attribute-sufficient bottleneck (Stage I) and conditional diffusion refinement (Stage II), preserves facial attributes and image quality to the greatest extent while ensuring extremely high anonymity (extremely low Re-ID), thus solving the problems of incomplete identity decoupling and low generation quality in existing technologies.

[0080] This invention not only achieves anonymization, but also, due to the latent space constructed in Stage I... It is highly interpretable and supports precise editing of specific facial attributes. During the inference phase, for the input image... Encoding yields latent vectors Assuming The Dimensions This corresponds to the attribute "whether or not wearing glasses". To add glasses to a character, simply modify the settings manually. By setting the value of the vector (e.g., making it positive) while keeping other dimensions unchanged, we obtain the modified vector. Then Decode and input the diffusion model.

[0081] Experimental results are as follows Figure 4 This demonstrates that, through the above operations, the generated anonymized image can accurately represent changes in the target attribute. For example, while keeping the person's posture, skin color, and hair color unchanged, operations such as "adding / removing glasses," "changing facial expressions (e.g., from not smiling to smiling)," and "changing hair color" were successfully implemented. During single-attribute editing, other non-target attributes remained unchanged. This proves that the attribute-sufficient bottleneck network of this invention successfully achieves decoupling between attributes, meaning that each latent dimension independently controls a semantic attribute.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A face anonymization generation method based on feature decoupling, characterized in that, Includes the following steps: Obtain the original face image to be anonymized; The original face image is input into the attribute-preserving bottleneck network. By applying attribute supervision and information bottleneck constraints in the latent space, the identity residual information in the image is filtered out, and the bottleneck image that preserves facial attributes but does not contain identity details is decoded. The bottleneck image is used as a conditional guiding signal and input into a pre-trained conditional diffusion refinement module. The random sampling generation mechanism of the diffusion model is used to restore the high-frequency texture and visual details of the image and generate the final anonymized face image.

2. The face anonymization generation method based on feature decoupling according to claim 1, characterized in that, The attribute-preserving bottleneck network is based on a variational autoencoder. The architecture is constructed, and its training and inference logic includes: (1) Encode the input image into a low-dimensional latent representation; (2) Attribute decoupling mechanism: Establish a one-to-one correspondence between each dimension of the potential representation and the predefined set of facial attributes, and supervise the training of the potential representation through the attribute classifier to ensure that each dimension encodes only specific attribute information; (3) Identity filtering mechanism: KL divergence is introduced as a regularization term to limit the information capacity of potential channels; by adjusting the regularization strength, the network is forced to encode only supervised attribute information, while identity texture features not included in the attribute set are discarded as redundant information. (4) The decoder reconstructs a bottleneck image based on the latent representation, which retains the facial structure and attributes but blurs the identity features.

3. The face anonymization generation method based on feature decoupling according to claim 1, characterized in that, The conditional diffusion refinement module is built based on the conditional diffusion probability model, and its specific mechanism includes: (1) Conditional guidance: The bottleneck image generated by the attribute-preserving bottleneck network is used as a conditional signal and input into the denoising network of the diffusion model; (2) Random generation: Using the reverse denoising process of the diffusion model, starting from pure Gaussian noise, the image is gradually restored under the guidance of the bottleneck image.

4. The face anonymization generation method based on feature decoupling according to claim 3, characterized in that, The diffusion model learns the statistical distribution of the data to randomly fill in reasonable high-frequency details.

5. The face anonymization generation method based on feature decoupling according to claim 2, characterized in that, Facial attributes can be edited, specifically by users directly modifying the dimensional values ​​corresponding to specific attributes in the latent representation; The modified latent representation is decoded into a new bottleneck image and input into the conditional diffusion refinement module to generate anonymized face images with changes in target attributes, thereby achieving precise control over face attributes or balancing of dataset distribution.

6. The face anonymization generation method based on feature decoupling according to claim 1 or 2, characterized in that, The total loss function for training a bottleneck network while preserving attributes for: in, To reconstruct the loss, the L1 distance is used; For binary cross-entropy loss, The weighting coefficients for attribute supervision loss are used to adjust the proportion of attribute classification in the total loss. Let KL divergence be the KL divergence. is the regularization weight coefficient of the KL divergence, used to control the constraint strength of the information bottleneck to adjust the degree of decoupling in the latent space.

7. The face anonymization generation method based on feature decoupling according to claim 1 or 3, characterized in that, Denoising training loss function for diffusion model Represented as: in, Indicates the original image The time step of random sampling and Gaussian noise Find the expected value of the joint distribution; Indicates random noise. This represents the noise component predicted by the denoising network. For noisy images, For time steps, For conditional images.

8. A face anonymization generation system based on feature decoupling, characterized in that, include: The feature extraction module is used to acquire the original face image and extract its multi-dimensional features; The feature decoupling and replacement module decouples multi-dimensional features into independent identity features and non-identity features through the attribute-preserving bottleneck network, filters out identity residual information in the image, and decodes to obtain a bottleneck image that preserves facial attributes but does not contain identity details; The image reconstruction module, through the conditional diffusion refinement module, generates anonymized face images based on the bottleneck image combination.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images